RESEARCHarXiv CS.LG·27d ago
Structural Interpretations of Protein Language Model Representations via Differentiable Graph Partitioning
This research proposes a framework to interpret protein language model representations by projecting them onto protein contact graphs and applying SoftBlobGIN, a Graph Isomorphism Network. This method performs structure-aware message passing to learn functional substructures, achieving 92.8% accuracy in enzyme classification and providing auditable structural explanations.
27